Practical and theoretical aspects of the sequential probability ratio test for anomaly detection

In this paper a method based on an autoregressive (AR) model using the sequential probability ratio test (SPRT) for anomaly detection is discussed. The method is compared theoretically and practically with a more conventional method, resulting in the SPRT method being superior to the last mentioned. The SPRT method is applied to thermocouple noise signals and to residual noise signals from an AR model that was established from the thermocouple noise signals. Results of anomaly detection show that the SPRT method is able to detect boiling in an early stage. Applying the SPRT method to residual noise is to be preferred to applying it directly to the thermocouple noise signals since it gives a better detection of anomalies with fewer false alarms. The AR-SPRT method is very well suited for anomaly detection. (authors)